Everything That Charged for a Single NLP Task Is Already Dead
In 2019, MonkeyLearn charged $299 a month so you could paste text into a box and get a sentiment score back. The box has since closed. The product technically still exists, somewhere in an acquisition. But what it was selling — a moat around a single text-classification call — has been undercut to a price that rounds to zero.
This isn't a story about one startup. It's a story about an entire category of software, and the founders who built it still haven't told their investors what the gap in their projections actually means.
The Narrow NLP API Had a Good Run
For about a decade, there was real money in wrapping a specific natural language capability behind an API and charging a subscription for access. Sentiment analysis. Named entity recognition. Topic classification. Language detection. Intent extraction. Document summarization. Translation.
These weren't trivial engineering problems in 2015. Building a reliable sentiment classifier that handled sarcasm, negation, and domain-specific vocabulary was months of training data, ML expertise, and infrastructure. If you needed it as a component in your product and you weren't an ML shop, you paid for access.
The market responded accordingly. AWS Comprehend. Google Cloud Natural Language API. IBM Watson NLP. Dozens of startups: Repustate, Lexalytics, ParallelDots, aylien, Meaning Cloud. Each one owned a niche. Each one had customers paying recurring revenue for access to capabilities their own teams couldn't build.
What Happened in 36 Months
GPT-3 arrived in 2020. By 2022, GPT-3.5 was accessible for essentially nothing per call. By 2023, GPT-4 had eliminated any performance argument for a specialized sentiment classifier. By 2025, you could run a capable open-source model locally, or pay OpenAI less than $0.001 per 1,000 tokens for their fastest model.
The math is brutal. AWS Comprehend charges $0.0001 per text unit for sentiment detection. That sounds cheap. But GPT-4o mini handles sentiment analysis, entity extraction, topic classification, intent detection, summarization, and translation — in any language, including ones Comprehend doesn't support, with context-awareness Comprehend never had — for $0.00015 per 1,000 tokens.
You're not choosing between specialized and cheap. You're choosing between specialized and more capable plus cheaper.
This is not normal competitive pricing pressure. This is a capability inversion.
The Graveyard Is Growing
MonkeyLearn was acquired by Medallia in 2022 and has effectively been wound down as a standalone product. Aylien (news NLP) sold its technology to Quantexa. Repustate pivoted. Others are zombies — still taking payments, no longer investing in the product, waiting to see if the customers notice.
The survivors either moved upstream fast enough, or they serve regulated industries where using a general-purpose LLM creates compliance complexity (healthcare HIPAA data, financial regulatory requirements, government contracts with data residency mandates). That carve-out is real but it's smaller than most of the incumbents need it to be.
AWS Comprehend's pricing page hasn't changed much, but look at their case studies. They're not winning new customers with "best-in-class sentiment analysis." They're winning on integration with existing AWS infrastructure, compliance certifications, and procurement inertia in enterprises already committed to the AWS stack.
That's not a product advantage. That's friction. And friction erodes.
The Mistake Founders Are Still Making
The product that dies is the one that says: "Give us text, we'll give you a label."
The product that survives is the one that says: "We've already analyzed ten million reviews in your industry, and here's what companies at your scale are missing."
The value was never in the classification call. It was in the aggregation, the benchmarking, the domain expertise layered on top of the call, the workflow integration that made the output actionable. Every startup that built a thin API wrapper over NLP capability and called it a product skipped the part that makes a product defensible.
An LLM can classify sentiment. It cannot tell you that your product's return rate is 40% higher in the Southeast because your sizing guide uses UK measurements, and three of your competitors have fixed this problem and gained 8 points of NPS in the same region. That analysis requires data you've collected, domain expertise you've developed, and a workflow your customers have built around your output.
That's the business. The classification call was always just plumbing.
Where the Value Actually Went
It didn't disappear. It compressed and moved upstream.
The companies capturing value now are building two layers above the inference call. Not "here's a sentiment score" but "here's what to do with the sentiment across your 50,000 reviews, segmented by your customer cohorts, ranked by business impact, with a suggested priority queue for your product team." Not a sentiment API. A product intelligence platform that happens to use sentiment analysis internally.
This matters because a lot of what's being built right now as "AI products" is still sentiment-API-shaped — thin wrappers over LLM calls charging SaaS prices for access to a capability that's becoming table stakes. Those products have a shorter runway than their founders think.
The test: if a technical founder at a mid-sized company could replicate your core output with three hours and a GPT-4 API key, you're not two layers above the inference call. You're one. And one isn't enough anymore.
The Surviving Pattern
The companies I see holding are the ones that recognized early that the inference layer was going to commoditize and built their moat somewhere else entirely:
Data moat. You've collected or licensed data that the LLM doesn't have. Historical records. Proprietary signals. Industry-specific training sets. The inference is cheap. The data is expensive.
Workflow integration. The LLM call is embedded in a workflow so deeply that replacing it means replacing the whole workflow. Salesforce, Notion, Figma are doing this — not selling AI, selling integrated tools that use AI internally.
Domain expertise operationalized. You've built the judgment layer. The model can analyze a legal contract for risk, but it can't prioritize which risks to negotiate versus accept for a Series B SaaS company versus a manufacturing firm. That prioritization is expertise. Expertise doesn't commoditize at the same rate as inference.
The narrow NLP API never had any of these. It was, always, a capability gap dressed up as a product. The gap closed.
The AI deployment theatre gap is real — but it's different from what's happening here. Deployment theatre is about companies claiming AI integration they haven't built. This is about companies that integrated correctly, but integrated into a layer that was always going to be eaten.
What you build next needs to be friction-proof — not against competitors, but against the underlying models getting better. Every quarter, the floor of what an LLM can do for $0.001 rises. Your product needs to be on the ceiling, not the floor.
The narrow NLP API had a good run. It's over.
Photo by Tima Miroshnichenko via Pexels.